Almost half of UK small businesses now use AI tools. Simply Business surveyed 1,842 small business owners this summer and found 47% using them, up from 22% a year earlier. The same survey asked what was holding firms back, and the top answer was security and privacy, named by 44%.
Both numbers come from the same survey of the same owners. The tools are in use, and the owners are not sure where the data goes.
The answer depends less on which AI you use than on which account your staff are signed into when they paste something in. That is good news, because the account is the cheap part to fix.
Most of it happens on personal accounts
Staff rarely wait for a company policy before trying a tool that saves them an hour. Microsoft and LinkedIn’s 2024 Work Trend Index surveyed 31,000 people across 31 countries. Of those using AI at work, 78% were bringing their own AI tools, and at small and medium-sized companies the figure was 80%.
In a small firm that often means a free or personal ChatGPT account in a browser tab, with a customer email or a spreadsheet of figures pasted into it. Nobody has done anything wrong. It is also the arrangement that gives the business the least control.
What the personal and business plans promise
OpenAI draws a clear line between its personal plans and its business ones.
On the personal plans (Free, Go, Plus and Pro), conversations can be used to train OpenAI’s models unless the person has switched off a setting called “Improve the model for everyone”. It sits under Settings, then Data controls. OpenAI says switching it off covers new conversations.
On the business plans the default runs the other way. In OpenAI’s words:
By default, we do not use your business data for training our models.
OpenAI will also sign a data processing addendum for ChatGPT Business, ChatGPT Enterprise and the API, which you request through a form. UK GDPR requires a written contract of this kind with any supplier that handles personal data on your behalf. A member of staff’s personal account does not give you one.
Microsoft and Google make similar commitments for work accounts. Microsoft says prompts and responses in Copilot “aren’t used to train foundation LLMs”. Google says it does not use your Workspace content to train generative AI models outside your organisation without permission.
So the same question, typed into the same tool, can sit under quite different terms depending on how the person signed in.
The Copilot risk is your own files
Microsoft’s paid Copilot licence, the version that works across your own files, settles the training question. It raises a different one.
It can search everything in your Microsoft 365 account that the person asking is allowed to see. Microsoft is plain about it:
Microsoft Copilot only surfaces organizational data to which individual users have at least view permissions.
That sounds safe until you remember how sharing in SharePoint and OneDrive tends to get set up. A folder shared with everyone in 2021 to get a job done is still shared with everyone. Nobody went looking in it, so it never mattered. Copilot goes looking. A salary spreadsheet in the wrong folder becomes the answer to a junior’s question about pay.
Nothing leaves the business in that case. It reaches the wrong person inside it. The fix is a permissions tidy before Copilot is switched on, not after.
Deleted does not always mean deleted
In May 2025 a US court ordered OpenAI to keep ChatGPT logs that would otherwise have been deleted, as evidence in the New York Times copyright case. OpenAI said the order covered the Free, Plus, Pro and Team plans (Team is now called Business) and the API, unless a customer had a zero data retention agreement. It did not cover Enterprise or Edu. The obligation ended for new conversations on 26 September 2025, and OpenAI has gone back to removing deleted chats within 30 days.
That may never happen again. What it shows is that a delete button describes what you see, and a court case your business is not part of can override it. So decide what goes into any of these tools on the assumption that it might be kept.
What going wrong looks like
The best-known case is Samsung’s. In spring 2023 its engineers pasted internal source code into ChatGPT, including code they wanted fixed, and one used it to turn a meeting recording into minutes. Samsung responded by temporarily banning generative AI tools on company devices and networks.
There was no hack. It was ordinary work, done in a tool the company had no agreement with. The realistic risk for a small firm looks the same: a helpful member of staff, and a customer’s details in the wrong account.
A ban is the instinctive response. It does not reach the phone in someone’s pocket, so it tends to move the use out of sight rather than stop it.
What to do this week
Ask what people already use. Ask openly, without blame. A ten-minute conversation will tell you more than any monitoring tool, and you need to know before you can set a rule.
Move regular users onto a business plan. If people use ChatGPT for work, pay for ChatGPT Business for them. If you already run Microsoft 365, Microsoft Copilot Chat comes with eligible business plans at no extra cost and may cover what they need. Either way the work then happens under business terms that say your data is not used for training.
Write a one-page rule. Say what never goes into any AI tool, whatever the account: passwords, bank and card details, anything covered by a confidentiality agreement, anything a customer told you in confidence. Say what is fine. Staff want to know where the line is, and a short rule they can remember beats a long one they cannot.
Check sharing before you switch on Copilot. Look for folders shared with everyone and files whose owner has left. An afternoon spent here is cheaper than the conversation after a pay spreadsheet turns up in an answer.
Treat personal data as you would with any supplier. If customer or staff details go into an AI tool, UK GDPR applies just as it does to your accounts package or your email provider. A business plan with a data processing addendum puts the AI tool on the same footing as those suppliers. A personal account does not.
When it needs more than a subscription
For most small firms the steps above are enough. They cost a subscription and an afternoon, and nothing has to be built.
Building something is worth it only when the question changes. Perhaps you want an assistant that answers only from your own documents and cites the page, or one that can read your systems without being able to change them. A general chat tool is the wrong shape for that however it is licensed. That is a narrower piece of work, and it should start from a clear use, not from a worry about data. We have written about the difference between a tool that helps and a system that finishes the work.
Frequently asked questions
Does ChatGPT use my data to train its models?
On the personal plans (Free, Go, Plus and Pro), conversations can be used for training unless the setting “Improve the model for everyone” is switched off under Data controls. On ChatGPT’s business plans and the API, OpenAI says data is not used for training by default unless the customer opts in.
Is it safe to put customer data into ChatGPT?
Not on a personal account. Customer details are personal data under UK GDPR, so they should only go to a supplier your business has an agreement with. OpenAI offers a data processing addendum for ChatGPT Business and Enterprise. Even then, decide what needs to go in and keep the rest out.
Is Microsoft Copilot safer than ChatGPT?
With a work account, Microsoft says Copilot’s prompts and responses are not used to train the underlying AI models, and it sits under your existing Microsoft agreement. The paid version’s particular risk is internal. It can surface any file the person asking already has permission to see, so careless SharePoint and OneDrive sharing becomes visible.
Should we ban ChatGPT at work?
A ban can move the use onto personal phones, where you have less control. A business account for regular users and a short written rule work better.
Flux Dynamics is a fractional CTO who builds. Most of the AI questions we are asked are answered by a subscription and a sensible rule rather than a project, and we will say so. Tell us what your team is using and we will give you a straight read.